Vehicle trip boundary intelligent judgment method fusing multi-dimensional features
By combining multi-dimensional feature fusion and machine learning models with DBSCAN clustering and dynamic parameter adjustment, the accuracy and adaptability issues in vehicle trip division are solved, achieving efficient and automated trip boundary identification and self-optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-07
AI Technical Summary
Existing technologies suffer from poor accuracy, weak adaptability, and low automation in vehicle trip segmentation. They fail to fully utilize the multi-dimensional features of trajectory data, leading to frequent misjudgments and omissions, and also incurring high deployment and maintenance costs.
A multi-dimensional feature fusion-based intelligent method for determining vehicle trip boundaries is adopted. Through data cleaning, feature extraction and modeling, combined with DBSCAN clustering and machine learning models, the method identifies stop points and integrates multiple rules to determine boundaries, dynamically adjusting parameters to adapt to different scenarios.
It significantly improves the accuracy and robustness of trip boundary determination, reduces human intervention, enhances adaptability, can handle boundary recognition in complex scenarios, and has continuous self-optimization capabilities.
Smart Images
Figure CN121808688A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent transportation, vehicle networking and big data analysis, and in particular to an intelligent method for determining vehicle trip boundaries by integrating multi-dimensional features. Background Technology
[0002] Accurate segmentation of vehicle trips is fundamental for applications such as fleet operation management, driving behavior analysis, logistics cost accounting, and traffic planning. Currently, mainstream trip segmentation methods primarily rely on preset rules or threshold judgments based on single-dimensional features. Their typical implementations and existing problems are as follows: Rule-based judgment methods based on fixed thresholds: These methods typically set a single or a few static thresholds as the basis for segmentation. While simple to implement, these methods have extremely poor adaptability. Different vehicle types, operating scenarios, and driver behavior habits vary significantly, and a single fixed threshold cannot be universally applied, resulting in low accuracy and frequent misjudgments and missed judgments in complex scenarios.
[0003] A simplified segmentation method based on stop point identification: This method identifies the start and end points of a trip by recognizing the locations where the vehicle remains stationary for extended periods. Common stop point detection algorithms include simple judgments based on speed and time thresholds. However, these methods have limited ability to distinguish between short-term, unintentional parking and long-term, intentional parking. When faced with complex urban road conditions, they are prone to generating a large number of invalid segments, and may miss segments when processing short-term business parking.
[0004] Segmentation models based on single features or simple logic: These methods may combine a few features such as time, distance, and speed, but typically use simple AND / OR logic for combination judgments. They fail to fully exploit the rich spatiotemporal pattern information contained in trajectory data, such as continuous changes in direction angle, time period characteristics of travel (day / night), acceleration patterns, etc. Therefore, their judgment ability is insufficient when facing scenarios with ambiguous boundaries, such as sudden changes in direction but short time intervals, or long periods of low-speed movement without actual stopping.
[0005] In summary, the existing technology has the following main drawbacks: Poor accuracy: In complex and ever-changing real-world operational scenarios, relying on fixed rules and simple logic makes it difficult to accurately capture the true boundaries of a trip, resulting in unreliable segmentation results.
[0006] Weak adaptability: The model parameters and rules are all statically preset and cannot be adaptively adjusted according to different vehicles, drivers, regions or seasons, resulting in high deployment and maintenance costs.
[0007] Low level of automation: The segmentation process requires a lot of manual intervention for parameter optimization and result correction, making it difficult to achieve large-scale, automated processing.
[0008] Insufficient utilization of features: The failure to systematically integrate the features of trajectory data in multiple dimensions such as time, space, direction, and motion state limits the potential for improving judgment accuracy.
[0009] Therefore, there is an urgent need for an intelligent method that can automatically, accurately, and adaptively divide vehicle trips to address the shortcomings of existing technologies in terms of accuracy, adaptability, and automation, and to meet the needs of modern smart logistics and refined fleet management. Summary of the Invention
[0010] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent method for determining vehicle trip boundaries that integrates multi-dimensional features.
[0011] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligently determining vehicle trip boundaries by integrating multi-dimensional features, comprising the following steps: S1: The data acquisition module collects raw vehicle trajectory data from the vehicle terminal and transmits it to the data cleaning module; the data cleaning module performs preprocessing to obtain a time-sorted sequence of trajectory points. Furthermore, the original vehicle trajectory data includes vehicle ID, timestamp ti, longitude ϕi, latitude λi, speed vi, heading angle θi, and status; The preprocessing includes: Standardize the format: unify the units and formats for time, coordinates, speed, etc. Quality filtering: Removes dirty data points with invalid latitude and longitude, abnormal speed, or incorrect timestamps; Deduplication and completion: Delete completely duplicate records and interpolate or fill in missing values for key fields. Sorting and Organization: For each vehicle's valid data, sort it in ascending order by timestamp to form structured data with the vehicle as the key and a time-ordered sequence of trajectory points as the values. Each point in this list contains core fields such as vehicle ID, timestamp, latitude and longitude, speed, and direction angle.
[0012] The trajectory point sequence is: P={p1,p2,…,pn}; where each trajectory point contains basic values: ti, ϕi, λi, vi, θi; S2: The feature fusion and modeling module receives the trajectory point sequence in S1 and extracts the multidimensional feature vector of each trajectory point in the trajectory point sequence; Furthermore, the feature fusion and modeling module transforms the trajectory point sequence p into a high-dimensional feature representation that can be used for decision-making. This module constructs a comprehensive feature vector by fusing multi-dimensional information such as time, space, direction, and motion state, and establishes a judgment model to provide a quantitative basis for trip boundary identification.
[0013] Extracting the multidimensional feature vector includes: Extract time-dimensional features, including time interval Δt i , cumulative driving time (Ticum), and nighttime driving ratio (Rinight); The formula for calculating the time interval feature is as follows:
[0014] The formula for calculating the cumulative driving time is:
[0015] The nighttime driving ratio Rinight is calculated based on the sunrise and sunset time function fnight(t), and its formula is as follows: ; Where ⨅() is an indicator function that returns 1 when it is nighttime and 0 otherwise; Extract spatial dimension features, calculate the spherical distance di between adjacent points and the cumulative mileage Dicum using the Haversine formula, and calculate the spatial displacements Δxi and Δyi. The formula for calculating the spherical distance di between adjacent points is as follows:
[0016] Where R = 6371 km, is the average radius of the Earth.
[0017] The formula for calculating the cumulative mileage (Dicum) is:
[0018] The formula for calculating spatial displacement characteristics is:
[0019] Extract directional dimension features and calculate heading angle variation Δθi, three-point curvature κi, and directional consistency σiθ; The formula for calculating the change in heading angle is: , The three-point curvature feature is calculated based on the previous point, the current point, and the next point, and its formula is:
[0020] Orientation consistency is characterized by the standard deviation of the heading angle within the sliding window, and the formula is:
[0021] Where θˉi is the arithmetic mean of the heading angles within the window.
[0022] Extract velocity dimension features, calculate velocity change rate Δvi, acceleration ai, sliding window velocity statistics μiv and σiv, and generate rapid acceleration / deceleration flags Iihard_acc and Iihard_dec; The formula for calculating the rate of change of velocity is: ; The formula for calculating acceleration is:
[0023] Where 3.6 is the conversion factor from km / h to m / s.
[0024] The formula for calculating the velocity statistics within the sliding window is: ; The formulas for calculating the acceleration and deceleration indicators are as follows:
[0025] Extract stop point features, apply the DBSCAN clustering algorithm to identify stop points and generate stop markers si; Feature fusion combines the features from the above dimensions into a unified feature vector fi and constructs a feature matrix; ; Where d is the feature dimension.
[0026] Feature standardization involves standardizing the feature matrix to obtain a standardized feature matrix F for use in subsequent models.
[0027]
[0028] S3: The boundary determination module identifies long-term stationary points and marks the trajectory sequence of the stationary points based on the multi-dimensional feature vector; Furthermore, the identification of long-term stationary points employs the density-based clustering algorithm DBSCAN, with the specific steps as follows: S31 defines a low-speed point set: From the trajectory point sequence P, trajectory points with speeds below a preset threshold v_threshold (2 km / h) are selected to form a low-speed point set P_low_speed = {p_i|v_i} <v_threshold}; S32: DBSCAN Clustering: Apply the DBSCAN algorithm to the low-speed point set P_low_speed. The algorithm parameters include the neighborhood radius ε and the minimum number of samples min_samples, and obtain the clustering result set C={c_1,c_2,...,c_m}, where each cluster c_k represents a potential dwelling region. S33 Stop point determination: For each cluster c_k, calculate its duration T_k=max_{p_i∈c_k}t_i-min_{p_i∈c_k}t_i. If T_k is greater than the preset minimum stay time T_min (10 minutes), then the cluster is determined to be a valid stop point. S34 Trajectory Point Dwell Marker: A dwell mark s_i is assigned to each trajectory point p_i, and its calculation formula is as follows:
[0029] S35 Stop Feature Enhancement: For each valid stop point, cluster c_k, and extract the stop duration T_k, the center point of the stop location (latitude φ_k^c=1 / |c_k|∑{p_i∈c_k}φ_i, longitude λ_k^c=1 / |c_k|∑{p_i∈c_k}λ_i), and the stop time period (start time t_k^start=min_{p_i∈c_k}t_i, end time t_k^end=max_{p_i∈c_k}t_i) as additional features, and incorporate them into the feature vector of the trajectory point sequence.
[0030] S4: The boundary determination module receives the trajectory sequence of stop points marked in S3. First, it performs a preliminary determination on each trajectory point based on multiple preset configurable rules and calculates the rule confidence score of each trajectory point as the trip boundary. Then, it inputs the multi-dimensional feature vector and the rule confidence score into the pre-trained machine learning model and outputs the probability value of each trajectory point as the trip boundary. Furthermore, the preset multiple configurable rules include: Rule 1: Abrupt change rule for heading angle. If the change in heading angle Δθ_i is greater than the preset threshold θ_th (120°), the rule is triggered. Rule 2: Time interval rule. If the time interval Δt_i is greater than the preset threshold t_th (16 hours), the rule is triggered. Rule 3: Nighttime Ratio Rule. If the nighttime driving ratio R_i^night is lower than the preset threshold r_th (0.5), the rule is triggered. Rule 4: Dwell time rule. If the dwell time marker s_i=1 and the dwell time is greater than the preset threshold T_stop (30 minutes), the rule is triggered. Rule 5: Speed drop rule. If the rate of change of speed Δv_i is less than the preset negative threshold v_drop (from normal driving speed to 0 km / h), the rule is triggered. Rule 6: Spatial distance mutation rule. If the spherical distance d_i between adjacent points is greater than the preset distance threshold d_th (10km), the rule is triggered.
[0031] Each rule is configured with an initial weight w_r (r=1,2,...,6), and the weight is adjustable. For each trajectory point p_i, the rule confidence score rule_score_i is calculated as the sum of the weights of all triggering rules, i.e.: ; Where R_i is the set of rules triggered at point p_i.
[0032] The pre-trained machine learning model employs the gradient boosting decision tree model XGBoost. Its input features include the multi-dimensional feature vector f_i extracted from S2 and the rule confidence score rule_score_i. The output is the probability value model_prob_i that the trajectory point p_i is a trip boundary. The model training process uses historical labeled data, where positive samples are manually labeled trip boundary points and negative samples are non-boundary points. The model hyperparameters are optimized through cross-validation and grid search.
[0033] S5: The boundary determination module integrates a dynamic parameter adjustment mechanism, which adaptively optimizes the weight allocation of the multiple rules and the determination parameters of the machine learning model based on the historical segmentation accuracy of the vehicle, manual correction feedback and real-time model evaluation indicators. Furthermore, the dynamic parameter adjustment mechanism includes: Rule weight adjustment: Based on historical segmentation accuracy, calculate the precision and recall of each rule, and dynamically adjust the rule weights according to the F1 score. The adjustment formula is as follows: ; Where α is the learning rate, β is the baseline F1 score, and F1_r is the F1 score of rule r; Model parameter tuning: Based on real-time model evaluation metrics (including accuracy, precision, recall, F1 score, and AUC) and human correction feedback, the machine learning model is subjected to online learning or periodic retraining to update the model parameters. Adaptive threshold adjustment: For the judgment thresholds in the rules (including θ_th, t_th, r_th, T_stop, v_drop, d_th), the thresholds are automatically adjusted using statistical methods (such as quantile analysis) based on vehicle type, operating area and historical behavior data to adapt to different scenarios.
[0034] S6: The post-processing module integrates the rule confidence score and machine learning probability value obtained in S4, determines the final trip boundary point sequence according to the set fusion strategy and threshold, and performs post-processing on the segmented results, such as merging short trips, removing anomalies, and smoothing boundaries. Furthermore, the fusion strategy employs a weighted linear fusion method to calculate the comprehensive score score_i for each trajectory point, using the following formula:
[0035] Where γ is the rule score weighting coefficient, with a value range of [0,1], which is adjustable.
[0036] The judgment threshold score_th (0.7) is set based on the comprehensive score. If score_i > score_th, then the judgment point p_i is the candidate trip boundary point.
[0037] The post-processing operations include: Merge short trips: If the travel time between two adjacent trip boundary points is less than the preset minimum trip duration T_min_trip (5 minutes), then merge these two boundary points, that is, remove the intermediate point from the boundary point sequence; Remove anomalies: If the mileage of a trip is less than the preset minimum mileage D_min_trip (0.5km), then the boundary point corresponding to that trip is removed. Smoothing Boundaries: Context checks are performed on boundary points. If the feature changes of the trajectory points before and after a boundary point are not significant (the difference between the comprehensive score score_i and the average score of the points before and after the boundary point is less than the threshold δ), the boundary point is considered to be removed to avoid misjudgment.
[0038] S7: The result output module formats the processed trip segment results for output, supports visualization and manual correction, and feeds the correction results back to the system as feedback data for continuous optimization of the model and parameters in S5.
[0039] Furthermore, the formatted output includes outputting the start and end times, start and end locations (latitude and longitude), mileage, travel time, average speed, nighttime travel ratio, and number of stops for each trip in a structured data format (including JSON, CSV, GeoJSON).
[0040] The visualization displays the original trajectory and trip segmentation results through a map interface. Different trips are distinguished by different colors or line types, and users can click to view detailed trip information.
[0041] The manual interactive correction allows users to add, delete, or move trip boundary points through a graphical interface, and the corrected results are stored as new annotation data in the annotation database.
[0042] The feedback data return refers to adding manually corrected labeled data to the historical labeled dataset to trigger online learning or periodic retraining of machine learning models and rule weights, thereby achieving continuous optimization of the system's judgment accuracy.
[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention significantly improves the accuracy and robustness of trip boundary determination. By deeply integrating multi-dimensional features such as time, space, direction, speed and stop points, and adopting a strategy of "multi-rule fusion and collaborative determination with machine learning models", it can more comprehensively and accurately identify trip boundaries in complex scenarios. It effectively overcomes misjudgment and omission caused by the limitations of a single rule or feature, and has a stronger tolerance to data noise and anomalies.
[0044] It achieves a high degree of adaptability and intelligence, reducing reliance on manual intervention. The system's built-in dynamic parameter adjustment mechanism can adaptively optimize rule weights, model parameters, and judgment thresholds based on vehicle type, operating area, historical behavior, and human feedback, eliminating the need for tedious manual reconfiguration for different application scenarios. Combined with an online learning and feedback loop, the system possesses continuous self-optimization capabilities, greatly improving its automation level and long-term applicability.
[0045] This invention enhances the ability to handle complex operational scenarios and ambiguous boundary conditions. Through density-based stop point clustering identification, targeted multi-rule design, and post-processing optimization, it can effectively distinguish between short-term stops and long-term stops, merge unreasonable short trips, and smooth boundary points, thereby maintaining accurate and reasonable segmentation results even in complex real-world scenarios such as frequent starts and stops and long-term low-speed driving.
[0046] This approach enhances the system's scalability and practical value. The modular design supports flexible upgrades and replacements of algorithm components, while standardized data interfaces facilitate integration with multi-source vehicle terminals and external management systems. This method provides efficient and reliable technical support for applications such as refined fleet management, logistics cost accounting, and driving behavior analysis, and has broad prospects for industrial application. Attached Figure Description
[0047] Figure 1 This is a flowchart illustrating the implementation of Embodiment 1 of the present invention. Detailed Implementation
[0048] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.
[0049] S1: The data acquisition module collects raw vehicle trajectory data from the vehicle terminal and transmits it to the data cleaning module; the data cleaning module performs preprocessing to obtain a time-sorted sequence of trajectory points. Furthermore, the original vehicle trajectory data includes vehicle ID, timestamp ti, longitude ϕi, latitude λi, speed vi, heading angle θi, and status; The preprocessing includes: Standardize the format: unify the units and formats for time, coordinates, speed, etc. Quality filtering: Removes dirty data points with invalid latitude and longitude, abnormal speed, or incorrect timestamps; Deduplication and completion: Delete completely duplicate records and interpolate or fill in missing values for key fields. Sorting and Organization: For each vehicle's valid data, sort it in ascending order by timestamp to form structured data with the vehicle as the key and a time-ordered sequence of trajectory points as the values. Each point in this list contains core fields such as vehicle ID, timestamp, latitude and longitude, speed, and direction angle.
[0050] The trajectory point sequence is: P={p1,p2,…,pn}; where each trajectory point contains basic values: ti, ϕi, λi, vi, θi; S2: The feature fusion and modeling module receives the trajectory point sequence in S1 and extracts the multidimensional feature vector of each trajectory point in the trajectory point sequence; Furthermore, the feature fusion and modeling module transforms the trajectory point sequence p into a high-dimensional feature representation that can be used for decision-making. This module constructs a comprehensive feature vector by fusing multi-dimensional information such as time, space, direction, and motion state, and establishes a judgment model to provide a quantitative basis for trip boundary identification.
[0051] Extracting the multidimensional feature vector includes: Extract time-dimensional features, including time interval Δt i , cumulative driving time (Ticum), and nighttime driving ratio (Rinight); The formula for calculating the time interval feature is as follows:
[0052] The formula for calculating the cumulative driving time is:
[0053] The nighttime driving ratio Rinight is calculated based on the sunrise and sunset time function fnight(t), and its formula is as follows: ; Where ⨅() is an indicator function that returns 1 when it is nighttime and 0 otherwise; Extract spatial dimension features, calculate the spherical distance di between adjacent points and the cumulative mileage Dicum using the Haversine formula, and calculate the spatial displacements Δxi and Δyi. The formula for calculating the spherical distance di between adjacent points is as follows:
[0054] Where R = 6371 km, is the average radius of the Earth.
[0055] Total mileage D i cum The calculation formula is:
[0056] The formula for calculating spatial displacement characteristics is:
[0057] Extract directional dimension features and calculate the heading angle change Δθ i Three-point curvature κ i and directional consistency σ i θ ; The formula for calculating the change in heading angle is: , The three-point curvature feature is calculated based on the previous point, the current point, and the next point, and its formula is:
[0058] Orientation consistency is characterized by the standard deviation of the heading angle within the sliding window, and the formula is:
[0059] Where θˉi is the arithmetic mean of the heading angles within the window.
[0060] Extract velocity dimension features, calculate velocity change rate Δvi, acceleration ai, sliding window velocity statistics μiv and σiv, and generate rapid acceleration / deceleration flags Iihard_acc and Iihard_dec; The formula for calculating the rate of change of velocity is: ; The formula for calculating acceleration is:
[0061] Where 3.6 is the conversion factor from km / h to m / s.
[0062] The formula for calculating the velocity statistics within the sliding window is: ; The formulas for calculating the acceleration and deceleration indicators are as follows:
[0063] Extract stop point features, apply the DBSCAN clustering algorithm to identify stop points and generate stop markers si; Feature fusion combines the features from the above dimensions into a unified feature vector fi and constructs a feature matrix; ; Where d is the feature dimension.
[0064] Feature standardization involves standardizing the feature matrix to obtain a standardized feature matrix F for use in subsequent models.
[0065]
[0066] S3: The boundary determination module identifies long-term stationary points and marks the trajectory sequence of the stationary points based on the multi-dimensional feature vector; Furthermore, the identification of long-term stationary points employs the density-based clustering algorithm DBSCAN, with the specific steps as follows: S31 defines a low-speed point set: From the trajectory point sequence P, trajectory points with speeds below a preset threshold v_threshold (2 km / h) are selected to form a low-speed point set P_low_speed = {p_i|v_i} <v_threshold}; S32: DBSCAN Clustering: Apply the DBSCAN algorithm to the low-speed point set P_low_speed. The algorithm parameters include the neighborhood radius ε and the minimum number of samples min_samples, and obtain the clustering result set C={c_1,c_2,...,c_m}, where each cluster c_k represents a potential dwelling region. S33 Stop point determination: For each cluster c_k, calculate its duration T_k=max_{p_i∈c_k}t_i-min_{p_i∈c_k}t_i. If T_k is greater than the preset minimum stay time T_min (10 minutes), then the cluster is determined to be a valid stop point. S34 Trajectory Point Dwell Marker: A dwell mark s_i is assigned to each trajectory point p_i, and its calculation formula is as follows:
[0067] S35 Stop Feature Enhancement: For each valid stop point, cluster c_k, and extract the stop duration T_k, the center point of the stop location (latitude φ_k^c=1 / |c_k|∑{p_i∈c_k}φ_i, longitude λ_k^c=1 / |c_k|∑{p_i∈c_k}λ_i), and the stop time period (start time t_k^start=min_{p_i∈c_k}t_i, end time t_k^end=max_{p_i∈c_k}t_i) as additional features, and incorporate them into the feature vector of the trajectory point sequence.
[0068] S4: The boundary determination module receives the trajectory sequence of stop points marked in S3. First, it performs a preliminary determination on each trajectory point based on multiple preset configurable rules and calculates the rule confidence score of each trajectory point as the trip boundary. Then, it inputs the multi-dimensional feature vector and the rule confidence score into the pre-trained machine learning model and outputs the probability value of each trajectory point as the trip boundary. Furthermore, the preset multiple configurable rules include: Rule 1: Abrupt change rule for heading angle. If the change in heading angle Δθ_i is greater than the preset threshold θ_th (120°), the rule is triggered. Rule 2: Time interval rule. If the time interval Δt_i is greater than the preset threshold t_th (16 hours), the rule is triggered. Rule 3: Nighttime Ratio Rule. If the nighttime driving ratio R_i^night is lower than the preset threshold r_th (0.5), the rule is triggered. Rule 4: Dwell time rule. If the dwell time marker s_i=1 and the dwell time is greater than the preset threshold T_stop (30 minutes), the rule is triggered. Rule 5: Speed drop rule. If the rate of change of speed Δv_i is less than the preset negative threshold v_drop (from normal driving speed to 0 km / h), the rule is triggered. Rule 6: Spatial distance mutation rule. If the spherical distance d_i between adjacent points is greater than the preset distance threshold d_th (10km), the rule is triggered.
[0069] Each rule is configured with an initial weight w_r (r=1,2,...,6), and the weight is adjustable. For each trajectory point p_i, the rule confidence score rule_score_i is calculated as the sum of the weights of all triggering rules, i.e.: ; Where R_i is the set of rules triggered at point p_i.
[0070] The pre-trained machine learning model employs the gradient boosting decision tree model XGBoost. Its input features include the multi-dimensional feature vector f_i extracted from S2 and the rule confidence score rule_score_i. The output is the probability value model_prob_i that the trajectory point p_i is a trip boundary. The model training process uses historical labeled data, where positive samples are manually labeled trip boundary points and negative samples are non-boundary points. The model hyperparameters are optimized through cross-validation and grid search.
[0071] S5: The boundary determination module integrates a dynamic parameter adjustment mechanism, which adaptively optimizes the weight allocation of the multiple rules and the determination parameters of the machine learning model based on the historical segmentation accuracy of the vehicle, manual correction feedback and real-time model evaluation indicators. Furthermore, the dynamic parameter adjustment mechanism includes: Rule weight adjustment: Based on historical segmentation accuracy, calculate the precision and recall of each rule, and dynamically adjust the rule weights according to the F1 score. The adjustment formula is as follows: ; Where α is the learning rate, β is the baseline F1 score, and F1_r is the F1 score of rule r; Model parameter tuning: Based on real-time model evaluation metrics (including accuracy, precision, recall, F1 score, and AUC) and human correction feedback, the machine learning model is subjected to online learning or periodic retraining to update the model parameters. Adaptive threshold adjustment: For the judgment thresholds in the rules (including θ_th, t_th, r_th, T_stop, v_drop, d_th), the thresholds are automatically adjusted using statistical methods (such as quantile analysis) based on vehicle type, operating area and historical behavior data to adapt to different scenarios.
[0072] S6: The post-processing module integrates the rule confidence score and machine learning probability value obtained in S4, determines the final trip boundary point sequence according to the set fusion strategy and threshold, and performs post-processing on the segmented results, such as merging short trips, removing anomalies, and smoothing boundaries. Furthermore, the fusion strategy employs a weighted linear fusion method to calculate the comprehensive score score_i for each trajectory point, using the following formula:
[0073] Where γ is the rule score weighting coefficient, with a value range of [0,1], which is adjustable.
[0074] The judgment threshold score_th (0.7) is set based on the comprehensive score. If score_i > score_th, then the judgment point p_i is the candidate trip boundary point.
[0075] The post-processing operations include: Merge short trips: If the travel time between two adjacent trip boundary points is less than the preset minimum trip duration T_min_trip (5 minutes), then merge these two boundary points, that is, remove the intermediate point from the boundary point sequence; Remove anomalies: If the mileage of a trip is less than the preset minimum mileage D_min_trip (0.5km), then the boundary point corresponding to that trip is removed. Smoothing Boundaries: Context checks are performed on boundary points. If the feature changes of the trajectory points before and after a boundary point are not significant (the difference between the comprehensive score score_i and the average score of the points before and after the boundary point is less than the threshold δ), the boundary point is considered to be removed to avoid misjudgment.
[0076] S7: The result output module formats the processed trip segment results for output, supports visualization and manual correction, and feeds the correction results back to the system as feedback data for continuous optimization of the model and parameters in S5.
[0077] Furthermore, the formatted output includes outputting the start and end times, start and end locations (latitude and longitude), mileage, travel time, average speed, nighttime travel ratio, and number of stops for each trip in a structured data format (including JSON, CSV, GeoJSON).
[0078] The visualization displays the original trajectory and trip segmentation results through a map interface. Different trips are distinguished by different colors or line types, and users can click to view detailed trip information.
[0079] The manual interactive correction allows users to add, delete, or move trip boundary points through a graphical interface, and the corrected results are stored as new annotation data in the annotation database.
[0080] The feedback data return refers to adding manually corrected labeled data to the historical labeled dataset to trigger online learning or periodic retraining of machine learning models and rule weights, thereby achieving continuous optimization of the system's judgment accuracy.
[0081] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.
Claims
1. A method for intelligently determining vehicle trip boundaries by integrating multi-dimensional features, characterized in that, Includes the following steps: S1: The data acquisition module collects raw vehicle trajectory data from the vehicle terminal and transmits it to the data cleaning module; the data cleaning module performs preprocessing to obtain a time-sorted sequence of trajectory points. S2: The feature fusion and modeling module receives the trajectory point sequence in S1 and extracts the multidimensional feature vector of each trajectory point in the trajectory point sequence; S3: The boundary determination module identifies long-term stationary points based on the multi-dimensional feature vector and marks the trajectory sequence of the stationary points; S4: The boundary determination module receives the trajectory sequence of stop points marked in S3. First, it performs a preliminary determination on each trajectory point based on multiple preset configurable rules and calculates the rule confidence score of each trajectory point as the trip boundary. Then, it inputs the multi-dimensional feature vector and the rule confidence score into the pre-trained machine learning model and outputs the probability value of each trajectory point as the trip boundary. S5: The boundary determination module integrates a dynamic parameter adjustment mechanism, which adaptively optimizes the weight allocation of the multiple rules and the determination parameters of the machine learning model based on the historical segmentation accuracy of the vehicle, manual correction feedback and real-time model evaluation indicators. S6: The post-processing module integrates the rule confidence score and machine learning probability value obtained in S4, determines the final trip boundary point sequence according to the set fusion strategy and threshold, and performs post-processing on the segmented results, such as merging short trips, removing anomalies, and smoothing boundaries. S7: The result output module formats the processed trip segment results for output, supports visualization and manual correction, and feeds the correction results back to the system as feedback data for continuous optimization of the model and parameters in S5.
2. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: In S3, the identification of long-term stationary points adopts the density-based clustering algorithm DBSCAN, which includes: defining a set of low-speed points; applying the DBSCAN algorithm to the set of low-speed points to obtain a set of clustering results; for each cluster, calculating its duration, and if it is greater than the preset minimum dwell time, it is determined to be a valid dwell point; assigning a dwell mark to each trajectory point; and extracting the dwell time, the center point of the dwell position, and the dwell time period as additional features.
3. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: In S4, the pre-trained machine learning model is the gradient boosting decision tree model XGBoost.
4. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: In S5, the dynamic parameter adjustment mechanism includes: dynamically adjusting rule weights based on historical segmentation accuracy; performing online learning or periodic retraining of the machine learning model based on real-time model evaluation indicators and manual correction feedback; and automatically adjusting the judgment threshold in the rules using statistical methods based on vehicle type, operating area, and historical behavior data.
5. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: In S7, the visualization displays the original trajectory and trip segmentation results through a map interface. Different trips are distinguished by different colors or line types, and clicks can be used to view detailed trip information. The manual interaction correction allows users to add, delete, or move trip boundary points through a graphical interface.
6. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: The extraction of multidimensional feature vectors in S2 includes: Extract time-dimensional features, including time intervals, cumulative driving time, and nighttime driving ratio; Extract spatial dimension features, including the distance between adjacent points on the sphere, cumulative mileage, and spatial displacement; Extract directional dimension features, including heading angle variation, three-point curvature, and directional consistency; Extract velocity dimension features, including velocity change rate, acceleration, sliding window velocity statistics, and rapid acceleration / deceleration indicators; Extract features of dwell points; The features from the above dimensions are combined into a unified feature vector, a feature matrix is constructed, and then standardized.
7. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 6, characterized in that: The nighttime driving ratio is calculated based on a sunrise and sunset time function; the spherical distance between adjacent points is calculated using the Haversine formula; the heading angle variation takes into account the periodicity of the angle; and the acceleration calculation includes the conversion of speed units from km / h to m / s.
8. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: In S6, the fusion strategy uses a weighted linear fusion method to calculate the comprehensive score of each trajectory point. The post-processing operations include merging short trips, removing abnormal trips, and smoothing boundary points.
9. The intelligent vehicle trip boundary determination method integrating multi-dimensional features as described in claim 1, characterized in that: The method also includes the automatic identification and processing of abnormal situations such as GPS signal loss, prolonged stationary periods, and extreme speed changes.
10. The intelligent vehicle trip boundary determination system integrating multi-dimensional features as described in claim 1, characterized in that, include: The data acquisition module is used to collect raw vehicle trajectory data from the in-vehicle terminal; The data cleaning module, connected to the data acquisition module, is used to preprocess the raw vehicle trajectory data to obtain a time-sorted sequence of trajectory points. The feature fusion and modeling module is connected to the data cleaning module and is used to extract the multidimensional feature vector of each trajectory point in the trajectory point sequence; The boundary determination module is connected to the feature fusion and modeling module, and is used to identify the stop point based on the multi-dimensional feature vector, and to fuse rules and machine learning models to determine the probability of the trip boundary. The post-processing module, connected to the boundary determination module, is used to fuse the determination results and perform post-processing to output the final trip segmentation result; The results output module is connected to the post-processing module and is used to format and display the segmented results and receive manual correction feedback.